Improved initial guess for minimum energy path calculations

TL;DR

The paper introduces an IDPP-based method for generating accurate initial guesses of transition paths, significantly reducing computational effort and improving convergence in minimum energy path calculations.

Contribution

A novel IDPP-based approach for initial path estimation that enhances convergence and reduces computational cost in minimum energy path calculations.

Findings

Reduced DFT computational effort by 50% to 90%.

Significantly improved initial path accuracy over linear interpolation.

Enhanced parallel computation efficiency due to load balancing.

Abstract

A method is presented for generating a good initial guess of a transition path between given initial and final states of a system without evaluation of the energy. An objective function surface is constructed using an interpolation of pairwise distances at each discretization point along the path and the nudged elastic band method then used to find an optimal path on this image dependent pair potential (IDPP) surface. This provides an initial path for the more computationally intensive calculations of the true minimum energy path using some method of choice for evaluating the energy and atomic forces, for example by ab initio or density functional theory. The optimal path on the IDPP surface is significantly closer to the true minimum energy path than a linear interpolation of the Cartesian coordinates and, therefore, reduces the number of iterations needed to reach convergence and averts divergence in the electronic structure calculations when atoms are brought too close to each other in the initial path. The method is illustrated with three examples: (1) rotation of a methyl group in an ethane molecule, (2) an exchange of atoms in an island on a crystal surface, and (3) an exchange of two Si-atoms in amorphous silicon. In all three cases, the computational effort in finding the minimum energy path with DFT was reduced by a factor ranging from 50 % to an order of magnitude by using an IDPP path as the initial path. The time required for parallel computations was reduced even more because of load imbalance when linear interpolation of Cartesian coordinates was used.